TikTok Virality Analysis
Author: Jaejun (J.J) Lee
Project Period: July 15th 2024 - August 15th 2024
Background
The value of social media content is vivid today. We recognize the presence of informational and promotional
content not only on social media but also in our daily lives due to its application to content and digital marketing.
TikTok has rapidly emerged as one of the leading social media platforms, with over 1 billion monthly active users
worldwide as of 2023 (Statista).
As the presence of social media content exponentially increases, the competition among social media platforms has
become fierce. TikTok faces competitors including Instagram Reels and YouTube Shorts, which introduced similar
short-form video features, leveraging their existing user bases. TikTok’s brand identity of short-form video features
has been threatened by its competitors, which would eventually impact the company’s value proposition and user
engagement. Further, TikTok confronts additional challenges in regulatory scrutiny over data privacy and security
issues followed by moderating content to ensure a safe environment, impacting its brand reputation. As such, we
believe TikTok should reinforce its core competency of short-form video features by understanding the factors
behind the virality of the platform’s main contents. The virality of short-form videos is a key feature that attracts
content creators and viewers. Understanding the factors that influence the virality of TikTok videos is crucial for
content creators aiming to maximize their reach and engagement, as well as for marketers and the platform itself to
enhance user experience and engagement. The discoveries of this study could be further developed as a manual for
potential creators who could ultimately impact the user engagement level of the platform. Interpreting the impact of
multiple factors on the virality of TikTok content would be mutually beneficial for users and creators seeking
guidance for virality and the company itself, which would eventually drive enhanced engagement, user retention,
and platform growth. In terms of analysis, analyzing user reviews utilizing sentiment analysis could provide a
deeper understanding of user satisfaction and preferences. Examining the relationship between different sentiments
expressed and stated in user reviews and the corresponding ratings would lead to achieving valuable insights into the
factors that influence user engagement and satisfaction. Applicable insights from the resulting factors can
consolidate strategies for enhancing the TikTok app and addressing user concerns.
Regarding competition, with the rapid growth of user-generated content, TikTok faces significant challenges in
content moderation. The company moderates millions of videos uploaded daily. Human moderation alone is
insufficient and costly, especially in mitigating misinformation. Videos presenting claims often require closer
scrutiny than other videos expressing opinions, as they are more likely to breach terms of service. To address this,
we decided to apply text mining and predictive analytics to analyze video transcriptions and develop a
machine-learning model to distinguish between claims and opinions. This model targets to improve TikTok's
moderation process by prioritizing videos for human review based on their content, which would subsequently
increase efficiency and ensure a safer user experience. We aim to create a predictive system that applies techniques
like topic modeling, sentiment analysis, and random forest algorithms. This system will help TikTok streamline
content review, focus on potentially harmful videos, and maintain platform integrity. As TikTok continues as a major
player in the social media landscape, understanding how different video characteristics impact user engagement has
become a norm for content creators and marketers. A method of gaining these insights is segmenting videos utilizing
key features such as text length and video duration. Identifying distinct clusters of videos could lead to uncovering
patterns that could inform improved content strategies.
The final section of cluster analysis targets to segment TikTok videos to reveal unique engagement patterns. We seek
to identify groups of videos with similar characteristics employing clustering techniques. Acquired actionable
insights could lead to optimizing viewer engagement. Understanding these segments would assist content creators in
tailoring and producing videos that resonate with audience preferences, which would enhance the overall
engagement, retention, and satisfaction on the platform. Regarding the analysis, we apply Hierarchical Clustering
and K-means Clustering to categorize videos based on text length and video duration. These methods allow us to
discover natural groupings within the data, which can then be applied to support strategic decisions for content
creation. Further, by examining the characteristics of each cluster, we also expect to develop targeted
recommendations for improving content performance and maximizing audience reach.
Literature Review:
The virality of TikTok videos is a multifaceted phenomenon influenced by numerous factors, including account
verification status, video content categories, and video length. Analyzing the virality of TikTok is critical for
understanding contemporary digital communication, engagement patterns, and how content spreads quickly
throughout public discourse. Analyzing TikTok can further find the broader trends in media consumption, the impact
of digital influencers, and the evolving dynamics of online social interactions. This knowledge would be vital and
applicable for researchers, marketers, and policymakers aiming to engage effectively with digital natives and
leverage the platform's potential for outreach and influence.
Peña-Fernández et al. (2022) indicate that verified accounts on TikTok often enjoy higher trust and engagement due
to perceived authenticity. Verified status can impact user interactions and content reach significantly. However,
Zannettou et al. (2023) found that unverified content may sometimes receive more views because viewers perceive
the unverified account as more genuine and approachable. The dynamics between verification status and virality
deliver a complex interplay where authenticity and relatability play crucial roles in user engagement; driven by
video virality.
TikTok has a wide range of video categories, specifically entertainment, humor, and trending challenges category are
more likely to go viral. These categories often attract more likes and comments due to their engaging nature and the
platform’s algorithm that promotes popular trends(Lee et al.,2022). The use of hashtags, music, and effects also
significantly enhances a video's likelihood of virality (Peña-Fernández et al.,2022).
Optimizing video length also enhances the pace of video spread out. According to a study, shorter videos tend to
perform better on TikTok, videos in the 0-30 seconds range maximize views, as they align with the platform's
fast-paced nature and the audience's short attention span. Videos between 31-60 seconds also perform well but tend
to have slightly lower engagement compared to the shortest videos. Videos longer than 60 seconds are generally less
popular unless they offer highly engaging content (Zannettou et al.,2023).
Additionally, the characteristics of TikTok creators show significant contributions to the virality of their videos.
Previous researchers found that higher follower counts and engagement rates are significantly associated with video
virality, as these metrics boost visibility through TikTok's algorithm (Pan et al., 2021). Creators with a large
follower base have a built-in audience that can amplify the reach of their content, increasing the likelihood of it
going viral. Additionally, a high engagement rate, characterized by a significant number of likes, comments, and
shares relative to the follower count, signals to the TikTok algorithm that the content is engaging, further boosting its
visibility.
While the previous research provides valuable insights into the factors contributing to video virality, further research
is needed to explore the long-term effects of viral growth rates. Our study aims to investigate the impact of different
content styles on a more diverse dataset within TikTok’s user base to address the problem of lack of generalizability
due to their focus on specific types of content. Our research will not only rely on engagement metrics to capture the
depth of user interaction but also the qualitative nature to allow its broader applicability.
Part.1 Sentiment Analysis of TikTok App Review
Introduction
This section presents the findings from a sentiment analysis of TikTok app reviews. The primary objective is to
examine the relationship between different sentiments expressed in user reviews and the corresponding ratings.
Sentiment analysis helps to understand how users feel about various features and updates of the TikTok app,
providing valuable insights for developers and marketers.
Data
The TikTok App Review dataset comprises user reviews and comments for the TikTok app available on the Google
Play Store. It includes variables such as reviewId (unique identifier), userName, userImage (profile image), content
(review text), score (rating from 1 to 5), thumbsUpCount (number of likes), reviewCreatedVersion (app version
reviewed), at (review creation date and time), replyContent (content of replies), and repliedAt (reply date and time).
Research Question:
How do different sentiments influence users’ ratings?
Hypotheses:
Null Hypothesis (H0): There is no significant correlation between any of the sentiments and user ratings.
Alternative Hypothesis (H1): There is a significant correlation between at least one of the sentiments and
user ratings.
Methodology
The dataset comprised 3,465,866 content reviews from the TikTok app. Each review was analyzed using the NRC
Emotion Lexicon to categorize words into different sentiments. The sentiments included joy, positivity, anticipation,
surprise, trust, fear, sadness, anger, disgust, and negative. The correlation between these sentiments and the review
ratings was calculated to determine the impact of each sentiment on the overall rating.
The following steps were involved in the analysis:
1. Data Collection and Data Cleaning: Cleaning the data and extracting relevant features.
2. Exploratory Data Analysis (EDA):
This line chart shows the daily number of reviews for the TikTok app on the Google Play Store
from 2021 to 2024.Peaks in review counts are evident in early 2021 and mid-2022, indicating
periods of high user engagement or significant events that prompted users to leave reviews. After
mid-2022, there is a noticeable decline in the number of daily reviews, which stabilizes at a lower
level compared to the earlier period.
This scatter plot displays the average review scores for different versions of the TikTok app.The
average review score varies across different app versions.There is a noticeable increase in
average review scores from version 14 to version 16, reaching a peak around version
16.Following version 16, the average review scores start to decline, with a significant drop in the
latest versions.This pattern suggests that specific updates or changes in certain versions were
well-received by users, while others were less favorable.
3. Sentiment Analysis: Applying the NRC Emotion Lexicon to identify sentiments in the reviews.
This bar chart shows the count of different emotions expressed in the reviews.Positive emotions
such as "positive" and "joy" are the most frequently expressed sentiments in the reviews.Negative
emotions such as "fear," "sadness," "anger," and "disgust" have lower counts but are still
present.The high frequency of positive emotions suggests a generally favorable user sentiment
towards the TikTok app.The presence of negative emotions indicates areas where users might be
experiencing issues or dissatisfaction.
4. Correlation Analysis: Calculating the correlation between sentiments and review ratings.
This set of bar charts shows the distribution of review ratings for different emotions.Reviews
expressing positive emotions (e.g., "joy," "positive") tend to have higher ratings (4 and 5
stars).Reviews expressing negative emotions (e.g., "anger," "disgust," "sadness") are more likely to
have lower ratings (1 and 2 stars).This reinforces the correlation between the type of emotion
expressed in the review and the review rating, with positive emotions linked to higher satisfaction
and negative emotions linked to dissatisfaction.
5. Visualization: Creating visual representations of the data to illustrate the findings.
Results
The analysis revealed the following correlations between sentiments and review ratings:
All correlations were statistically significant with p-values less than 0.05. The results indicate that positive
sentiments such as joy and positive have a significant positive correlation with higher review ratings. Conversely,
negative sentiments such as fear, sadness, anger, disgust, and overall negative sentiment have a significant negative
correlation with review ratings. This suggests that users who express positive emotions in their reviews tend to give
higher ratings, while those expressing negative emotions tend to give lower ratings.
Conclusion
The sentiment analysis of TikTok app reviews provides valuable insights into user satisfaction and areas for
improvement. Positive sentiments are associated with higher ratings, indicating user satisfaction, while negative
sentiments correlate with lower ratings, highlighting potential issues or dissatisfaction. These findings can help
TikTok developers and marketers address user concerns and enhance the overall user experience.
Part 2. Text Mining and Prediction
This section focuses on the application of text mining and predictive analytics to analyze and forecast patterns in
textual data. The main objectives are to extract meaningful information from large text corpora and to develop
predictive models. We want to use machine learning techniques to predict a binary outcome variable.
The purpose of this model is to mitigate misinformation in videos on the TikTok platform. The goal of this model is
to predict whether a TikTok video presents a "claim" or presents an "opinion".
Purpose
TikTok users can report videos they believe breach the platform's terms of service. Given the millions of videos
created and viewed daily, an overwhelming number of reports are generated, making it impossible for human
moderators to review each one individually. Analysis reveals that when terms of service violations occur, the content
is more likely to present a claim rather than an opinion. Consequently, distinguishing between claims and opinions
in videos is beneficial.
The bar chart provides a comparative analysis of author ban statuses within two distinct claim
statuses: "claim" and "opinion." A notable difference is observed in the "claim" category, which
shows a higher count of authors under review and banned compared to the "opinion" category.
This suggests that claims are more scrutinized, leading to a greater proportion of authors being
under review or banned.
We aim to develop a machine learning model to identify claims and opinions. Videos identified as opinions would
have a lower priority for human moderation. In contrast, videos identified as claims would undergo further
processing to determine their review priority. For instance, claim-classified videos might be ranked by the frequency
of reports, with the top percentage reviewed by humans daily.
A machine learning model would significantly enhance the process of presenting human moderators with videos
most likely to violate TikTok's terms of service.
Methodology
The dataset comprised 3,465,866 reviews from the TikTok app. Each review was analyzed using the NRC Emotion
Lexicon to categorize words into different sentiments. The sentiments included joy, positive, anticipation, surprise,
trust, fear, sadness, anger, disgust, and negative. The correlation between these sentiments and the review ratings
was calculated to determine the impact of each sentiment on the overall rating.
The following steps were involved in the analysis:
1. Data Collection and Data Cleaning: Extracting video data, cleaning and preprocessing the data to ensure
consistency and accuracy. This includes removing stop words, stemming, and lemmatization. Extracting
relevant features from the text, such as keywords.
2. Exploratory Data Analysis (EDA): Visualizing the data to identify initial patterns and relationships.
The histogram reveals distinct patterns in the transcription lengths of videos categorized as
claims and opinions. Both distributions exhibit a roughly normal shape with a slight right skew,
indicating that most transcription lengths cluster around the mean. Notably, claim videos tend to
have longer transcriptions compared to opinion videos. The blue bars extending further to the
right and the peak for claims occurring between 80 and 110 characters, compared to the peak for
opinions at 70 to 90 characters. On average, claim transcriptions are about 13 characters longer
than those of opinions. This suggests that claim videos are generally more detailed or contain
more information, which could imply a greater need for elaboration in claims compared to
opinions.
3. Text Mining Techniques: Applying techniques like topic modeling, sentiment analysis, and keyword
extraction to gain deeper insights into the text data.
the chart highlights the most common words used in video transcriptions, reflecting themes of
communication, information, relationships, and global topics. This provides insight into the content focus
of the videos, suggesting a strong emphasis on claims, learning, media, and social interactions.
4. Predictive Modeling: Using random forest to build models that can predict the outcome.
The variable importance plot from the random forest model highlights the significance of various
features in predicting the target variable. The most influential variables are engagement
metrics.These features have the highest Mean Decrease Gini values, indicating their substantial
contribution to the model's predictive accuracy. Moderately important variables include
content-related terms such as claim, read, and media, along with social interaction indicators like
friend and colleague. Less critical variables, such as text_length, discussion, and forum, show
lower Mean Decrease Gini values, suggesting a minimal impact on the model's performance.
5. Model Performance: Using confusion matrix to elevate the model
The confusion matrix shows that the classification model performs exceptionally well in distinguishing between
"claim" and "opinion" categories. With an accuracy close to 100%, and very high precision and recall for both
categories, the model demonstrates robust predictive capability and minimal error rates. This indicates that the
model is highly reliable for the given classification task.
True Negatives (TN): 1900
False Positives (FP): 0
False Negatives (FN): 1
True Positives (TP): 1916
Conclusion
An ideal model would have only true negatives and true positives, with no false negatives or false positives. As
demonstrated in the confusion matrix, this model achieves zero false negatives. Generate a classification report
including precision, recall, f1-score, and accuracy metrics to assess the model's performance. This model
performed well on both the validation and test holdout data. Furthermore, both precision and F1 scores were
consistently high. The model very successfully classified claims and opinions.
Part 3: Cluster Analysis
Purpose
The objective of this study is to segment TikTok videos based on their text length and video duration to identify
distinct clusters that exhibit unique engagement patterns. By understanding these clusters, we aim to provide
actionable insights that can inform content strategy and optimize viewer engagement.
Methodology
1. Handling Missing Values
Missing values in the dataset can significantly impact the clustering results. We handled missing values by removing
records with missing values in critical fields. For other missing values, imputation was performed using the mean or
median of the respective attributes. To ensure all attributes contribute equally to the clustering process, we
standardized the numerical attributes using z-score normalization, which scales the data to have a mean of zero and a
standard deviation of one.
2. Clustering Methods
K-Means Clustering
K-Means clustering is a partitioning method that divides the dataset into clusters. The algorithm
iteratively assigns each data point to the nearest cluster centroid and recalculates the centroids until
convergence. To determine the optimal number of clusters, we used the Elbow Method, which involves
plotting the total within-cluster sum of squares (WCSS) against the number of clusters and identifying
the "elbow point" where the rate of decrease slows down.
Hierarchical Clustering
We employed both Hierarchical Clustering and K-means Clustering to identify meaningful segments
within the dataset. For Hierarchical Clustering, Euclidean distance was used to compute the
dissimilarity matrix, and Ward’s method was used to minimize the variance within clusters. The
dendrogram was analyzed to identify the optimal number of clusters, focusing on two and three
clusters. In K-means Clustering, the optimal number of clusters was determined using the Elbow
Method, with the final cluster formation performed using k=3.
Results
1) K-Means Visualization
Performing K-means clusterin, we focus on the text_length and video_duration_sec variables and shows the
following cluster characteristics. This plot visualizes the results of the K-means clustering performed on the TikTok
video dataset.
Cluster 1 (Red) Represents videos with relatively long text lengths and varying video durations. These videos are
likely more descriptive and detailed. The second Cluster 2 (Green) contains videos with moderate text lengths and
shorter video durations. These videos might strike a balance between brevity and informativeness. The Cluster 3
(Blue) encompasses videos with short text lengths and varying video durations. These are likely concise videos that
get straight to the point.
2) Importance Scores by Feature
We converted the summarized data to a long format suitable for visualization and created a plot to show the
importance scores of different features for each cluster. The plot illustrates how features such as text length, video
duration, and various engagement metrics contribute to the formation of distinct clusters.
For Cluster 1, represented by the red line, text length shows a high positive importance score, indicating that videos
with longer descriptions are significant in this cluster. Claim status also has a moderately high importance score,
suggesting its relevance. In contrast, features like video duration, like count, share count, and view count have low
or negative importance scores, implying they are less influential in defining this cluster.
Cluster 2, depicted by the green line, is characterized by videos with longer durations, as indicated by the high
positive importance score for video duration. This cluster also shows a significant positive importance score for
video like count, highlighting the importance of likes in distinguishing these videos. Video share count and view
count have positive but less pronounced importance scores. Interestingly, text length and author ban status have
negative importance scores, indicating these features are less relevant for Cluster 2.
Cluster 3, shown by the blue line, reveals a moderate positive importance score for claim status, indicating its
relevance for this cluster. However, text length and video duration have negative importance scores, suggesting these
features are less critical. Engagement metrics such as like count, share count, and view count have low importance
scores, indicating that these metrics do not significantly distinguish Cluster 3.
These findings suggest that different clusters of TikTok videos can be characterized by specific features. For Cluster
1, longer text descriptions are significant, whereas Cluster 2 is defined by longer video durations and higher like
counts. Cluster 3, on the other hand, is influenced by claim status with less emphasis on text length and video
duration.
Content creators can use these insights to tailor their videos for better engagement. For example, they might focus
on crafting longer descriptions for Cluster 1 or producing longer, engaging videos to maximize likes for Cluster 2.
Understanding the specific characteristics that define each cluster allows for more targeted content strategies,
potentially enhancing viewer engagement and visibility on TikTok. Further analysis incorporating additional features
and refining the clustering process will provide deeper insights and more robust recommendations.
Lastly, we extracted additional features from the video dataset for segmentation analysis. These features were scaled
to standardize them, and the mean value of each feature for each cluster was calculated to better understand the
characteristics of each cluster.
The line plot shows the importance scores of each feature for each cluster. The plot indicates which features are most
important for each cluster. For example, one cluster may have high importance scores for video_duration_sec and
text_length, suggesting that these features are key characteristics of that cluster. This helps in understanding the
distinguishing features of each cluster and can inform targeted strategies or interventions.
Discussion
The cluster analysis reveals three distinct segments within the TikTok video dataset. Each cluster exhibits unique
characteristics that can be leveraged to tailor content strategies effectively. Cluster 1 indicates a preference for longer
videos with concise text, suggesting an emphasis on visual engagement. Cluster 2 highlights a need for quick,
information-rich videos, aligning with users seeking fast yet detailed content. Cluster 3 represents a versatile
segment where a balanced approach can capture a wider audience.
Conclusions
The findings from this analysis provide valuable insights into the content preferences of TikTok users. By aligning
content strategies with the identified clusters, we can enhance viewer engagement and retention. Based on the cluster
characteristics, the following recommendations are proposed: For Cluster 1, focus on creating long, engaging videos
with captivating visuals and concise text to enhance viewer retention and encourage repeated viewing. For Cluster 2,
produce short, informative videos with detailed text to deliver value quickly, catering to viewers who prefer rapid
consumption of content. For Cluster 3, implement a balanced content strategy that combines elements of both
engagement and information to appeal to a diverse audience and maximize overall reach.
References
Lee, S., Kim, J., & Koo, D. (2022). The Impact of Video Content Categories on Virality in TikTok. Journal of
Digital Marketing, 15(3), 145-160.
Pan, X., Zhang, J., & Wang, Y. (2021). Characteristics of TikTok Creators and Their Impact on Video Virality.
Social Media Research Journal, 10(4), 298-312.
Peña-Fernández, A., Martín, S., & González, R. (2022). Trust and Engagement: The Role of Account Verification on
TikTok. International Journal of Social Media Studies, 9(2), 201-215.
Statista. (2023). Number of monthly active TikTok users worldwide from 2018 to 2023. Retrieved from
https://www.statista.com/statistics/1231466/tiktok-global-mau/
Zannettou, S., Bradshaw, S., & Howard, P. N. (2023). Unverified Content and Its Role in TikTok's Virality. New
Media & Society, 25(1), 35-54.